This example demonstrates how to use the rocProfiler SDK for asynchronous, device-level hardware counter collection. Unlike dispatch-specific profiling, this method samples counters periodically on a separate thread. The application configures a device counting service, which runs in the background and collects data from a specified GPU agent.

1. Tool Loading and Initialization:
   • The rocProfiler runtime loads the tool's shared library and calls the rocprofiler_configure entry point.
   • The tool registers its tool_init and tool_fini functions.
   • The tool_init function creates a rocProfiler context, a buffer for the counter data, and discovers the first available GPU agent.
2. Profile and Service Configuration:
   • A counter profile for SQ_WAVES is built and cached for the selected agent.
   • The rocprofiler_configure_device_counting_service is configured. It uses a set_profile callback to provide the counter configuration to the service.
3. Asynchronous Sampling:
   • A detached thread is launched, which starts the rocProfiler context.
   • Inside this thread, rocprofiler_sample_device_counting_service is called in a loop to trigger counter collection periodically (every 50ms).
4. Workload Execution:
   • Concurrently, the main function in main.cpp launches a series of HIP kernels (kernel_a, kernel_b, kernel_c) in a loop to generate GPU activity.
5. Data Collection and Processing:
   • The sampling thread collects the counter data and stores it in the buffer.
   • When the buffer's watermark is reached, the buffered_callback is invoked to process and print the counter values.
6. Finalization:
   • After the workload is complete, the tool_fini function signals the sampling thread to stop, stops the context, and flushes the buffer to ensure all data is processed.

• Device Counting Service:

  • rocprofiler_configure_device_counting_service(): Configures a service for profiling an entire GPU device asynchronously.
  • The service is configured for a single rocprofiler_agent_id_t, allowing for targeted profiling of one GPU.

• Asynchronous Sampling:

  • rocprofiler_sample_device_counting_service(): Explicitly triggers a counter collection sample. In this example, it is called from a dedicated thread to perform periodic, time-based monitoring, independent of kernel dispatches.

• Profile Callback:

  • A set_profile callback is used by the service to request the rocprofiler_counter_config_id_t for the targeted agent.

• Buffer and Callback:

  • A rocprofiler_buffer_id_t is used to store the sampled counter data, which is then processed by a buffered_callback.

• rocprofiler_assign_callback_thread
• rocprofiler_configure_device_counting_service
• rocprofiler_create_buffer
• rocprofiler_create_callback_thread
• rocprofiler_create_context
• rocprofiler_create_counter_config
• rocprofiler_flush_buffer
• rocprofiler_iterate_agent_supported_counters
• rocprofiler_query_available_agents
• rocprofiler_query_counter_info
• rocprofiler_sample_device_counting_service
• rocprofiler_start_context
• rocprofiler_stop_context

• hipDeviceReset
• hipDeviceSynchronize
• hipFree
• hipGetDeviceCount
• hipGetDeviceProperties
• hipLaunchKernelGGL
• hipMalloc
• hipMemcpy
• hipSetDevice

• rocprofiler_buffer_id_t
• rocprofiler_callback_thread_t
• rocprofiler_context_id_t
• rocprofiler_counter_config_id_t
• ROCPROFILER_BUFFER_CATEGORY_COUNTERS
• ROCPROFILER_BUFFER_POLICY_LOSSLESS
• ROCPROFILER_COUNTER_RECORD_VALUE